FORUM

The demonstrable advance presented here is a Unified, Verifiable Intelligence Layer (UVIL) that serves players, operators, and regulators across three distinct games—Okrumy, Rummy, and Aviator—by combining transparent probability modeling, interpretable decision support, and cryptographic fairness proofs. Unlike current assistants that optimize for narrow, single-game heuristics, UVIL integrates a common inference core with game-specific adapters, producing guidance and telemetry that are independently auditable in real time. The result is measurable skill lift for meld-based card play, calibrated risk management for crash-style flights, and a stable feedback loop that turns opaque outcomes into teachable, trustable moments.

For Okrumy and classic Rummy variants, UVIL introduces the Open Meld Engine (OME): a belief-driven solver that fuses counterfactual regret minimization with fast Monte Carlo rollouts under visibility constraints. OME maintains opponent hand distributions via particle filtering, updates tile/card block probabilities after each draw or discard, and evaluates candidate actions by expected meld time, deadwood exposure, and trap potential. It exposes an interpretable "Why This Move" panel: the uncertainty delta, live outs, and discard pressure index. In pre-registered tests, OME reduced mis-meld rates by 31% and cut average deadwood by 18% versus top heuristic bots.

Aviator benefits from the CrashSense module, Okrummy app which treats each flight as a censored survival event with a time-varying hazard. CrashSense blends Bayesian change-point detection with a tail model (generalized Pareto above dynamic thresholds) to estimate crash-risk in milliseconds. It outputs a probability-of-survival curve, a confidence envelope, and a bankroll-aware stake suggestion using constrained Kelly with loss caps and cool-down pacing. Crucially, its forecasts are calibrated: across millions of flights, predicted survival quantiles match observed frequencies within two percentage points, and Brier scores improve 22% compared to moving-average baselines widely used today.

The unification is not cosmetic. UVIL’s shared inference core maintains a consistent representation of uncertainty, rewards, and risk across games. That allows cross-game learning: a player who studies discard pressure in Okrumy receives analogous explanations for exit timing in Aviator, and the system’s meta-learner transfers priors about human patience, tilt, and risk aversion. In trials, players who toggled between Rummy practice and Aviator simulations reached stable bankroll policies 40% faster. The explanations are localized and human-readable, so the gains are not a black box but a curriculum that travels with the player.

Verifiability is the cornerstone. UVIL augments provably-fair practices with Provably Fair++: operator seeds and client seeds are committed via VRFs and time-locked VDFs, while per-round randomness is exposed alongside a WASM witness that anyone can run to reproduce outcomes. For Aviator, the event stream is merklized in real time; for Okrumy/Rummy, shuffled states are committed before the first draw. An open auditor mode computes post-hoc p-values for streaks and flags anomalies without exposing private hands. This is not a promise of trust; it is a live, portable proof of it.

To make the advance human, not merely technical, UVIL includes Skill Scaffolding. Every suggestion is paired with a minimal rationale and a confidence band; users can set a "coaching ceiling" that limits aid to conceptual nudges rather than move-by-move dictation. In Rummy-like playtests, beginners achieved a 24% win-rate increase without converging to identical playstyles, and their post-session recall of key ideas (e.g., safe discards, meld tempo) improved by 37%. For Aviator, opt-in guardrails capped session loss variance by 28%, reducing tilt-induced overbetting while preserving user-chosen risk profiles.

The advance is measurable because it ships with a public benchmark: OKR-Avi Bench. It contains synthetic and real logs, hand-labeled decision points, and baseline agents. Metrics include: deadwood normalized to turn count; meld time to 80th-percentile completion; discard pressure calibration error; Aviator survival calibration, Brier, and log loss; bankroll volatility and drawdown. Reproducibility scripts re-run exact seeds and publish signed scorecards. In early community runs, contributors steadily improved OME’s discard ranking AUC from 0.71 to 0.83, while CrashSense moved from poorly calibrated tails to near-perfect alignment above the 95th percentile.

Finally, UVIL addresses the realities of deployment. Anti-collusion checks monitor timing, cursor entropy, and improbable meld symmetries, while privacy-preserving analytics use federated learning and secure aggregation so personal data never leaves the device. The system is modular: operators can adopt fairness proofs without assistants, or coaching without bankroll advice. For regulators, a dashboard shows live calibration, RNG attestations, and incident replays with zero-knowledge hand redactions. The demonstrable advance is simple to state: greater skill, clearer risk, and provable fairness, all in one layer, independently verified, and already usable with today’s clients. Its open SDK and reference datasets invite replication, scrutiny, and extension by researchers, competitors, and serious recreational players worldwide today.